Collective Robotics

IAR Lecture 12

Barbara Webb

Collective robotics

Different approaches (Kernbach, 2013)

• Co-operative: distributed sensing and actuation, but

centralised control.

• Networked: higher individual autonomy, but still high

level of communication and common knowledge.

• Swarm: no common knowledge, only local

communication, or interaction via effects on the world.

• Small world: minimalist capabilities of individual,

collective computation.

E.g. exploring with multiple robots

• Provided robots can merge their maps, can explore faster

with multiple robots.

• Potential speed up is 2*k, as single robot would need to

spend time traversing known space to get to new frontier.

• But need to co-ordinate exploration.

Active mapping for k robots – For each robot, obtain cost function of moving from

current position to other possible grid positions.

– Use binary representation of potential information gain:

cell is 0 if explored, 1 if unexplored.

– For each robot in turn, set goal as unexplored grid

location that has minimum cost to reach, and mark that

location as explored (so not available for next robot).

• More sophisticated approach would allow robots to swap

goals if this reduces the overall cost, e.g. using auction

mechanism.

• Generalised, this is the problem of task allocation.

E.g. Behaviour based task allocation

• ALLIANCE architecture (Parker,1998) – Robots have motivation systems determining action

selection:

• Impatience: will choose a task not being completed

by other robots

• Acquiescence: give up a task if failing to complete

– Broadcast periodic messages to each other indicating

what they are doing.

– Sensory feedback to monitor progress on tasks.

Emergent co-operation

• Holland (1995) ‘Stigmergy’

• Robots:

– Front scoop tends to collect pucks

– Lever triggers switch if pushing two or more, makes robot back up, leaving pucks behind

– Also avoid walls and each other using IR.

• Result is gradual aggregation of pucks in a single pile

Emergent co-operation

• Melhuish et al (2000)

• Robots can carry puck, detect gradient and notice

if cross a boundary line.

• Simple rules:

– If hit another object, drop puck

– If cross boundary going up gradient, move short

distance and drop puck

– If cross boundary going down gradient, back-up short

distance and drop puck

Emergent group behaviour

Flocking: (Reynolds 1987)

Assumes all ‘boids’ are identical and

follow the same local rules:

1. Collision Avoidance: Separate

from other boids.

2. Centering: Stay close to other

boids.

3. Velocity matching: travel in

same direction.

Flocking in real robots is difficult:

• various attempts, but needed to include virtual or explicit

leader, or all robots sensing goal

• also problem of how to make individual robot able to sense

relative position and bearing of neighbours

• Recent example addresses some of these limitations (Turgut et

al., 2008)

Emergent group behaviour

Turgut et al. 2008

• IR sensors used to detect other robots (when not active)

and obstacles (when active)

• Use virtual heading sensor: each robot has a compass and

wirelessly broadcasts its direction to neighbouring robots.

• Desired heading alignment is calculated as average of detected neighbours

• Proximal control is calculated as a virtual force with respect to detected robots or obstacles

• Is robust to noise, but find coherent swarm size depends on virtual heading sensor range: noise in system prevents long-range order emerging from short-range interactions.

• Large flocks possible with just a few long-range interactions or with some common homing information.

Self organised construction

Kilobots (Rubenstein et al.2014)

Local communication to localise and navigate

Kilobots (Rubenstein et al.2014)

Kilobots (Rubenstein et al.2014)

• Basic task allocation assumes all

robots can do all tasks and just

need to distribute them effectively,

then work separately.

• More complex scenarios:

– Task might require complex

continuous interaction between

two or more robots.

– Robots could be heterogeneous.

– Robots could be interacting with

other technologies, or humans

(or animals).

Inverse flocking: modelling duck behaviour by simple

flocking rules to produce sheepdog control algorithm

(Vaughan et al 2000)

References Serge Kernbach (2013) Handbook of Collective Robotics. CRC Press. Lynne Parker (1998) ALLIANCE: An Architecture for Fault Tolerant Multi-

Robot Cooperation IEEE Transactions on Robotics and Autonomous Systems, 14(2):220-240

Owen Holland and Chris Melhuish (1999) Stigmergy, self-organization, and sorting in collective robotics Artificial Life 5:173 - 202

Craig W.Reynolds (1987) Flocks, Herds, and Schools: A Distributed Behavioral Model, in Computer Graphics, 21(4) (SIGGRAPH '87 Conference Proceedings) pages 25-34. (http://www.red3d.com/cwr/papers/1987/boids.html)

Turgut, A.E., Celikkanat, H., Gokce, F. and Sahin, E. (2008) Self organized flocking in mobile robot swarms. Swarm Intelligence 2: 97-120

Rubenstein, M., Cornejo, A., Nagpal, R. (2014) Programmable Self-Assembly in a Thousand-Robot Swarm. Science 345, no. 6198 pp. 795-799

Vaughan, R., Sumpter, N., Henderson, J., Frost, A. and Cameron, S. Experiments in Automatic Flock Control (2000) Journal of Robotics and Autonomous Systems 31 pp.109-117